Friday’s Party is ultimately about pre-summarized context to reduce load.
I disagree with pulling entity state or aliases out of the context (you lose more than you gain, except when you need NO live context for reasons.) so not a fan of dropping ents to save space it leads to erratic behavior… .Do they all need to be there no… Ninhave about 10% of mine. But added context should absolutely be RAG long term. But it doesn’t all HAVE TO BE RAG to reduce context load. Some can and should be dynamic prompting based on state data to change behavior, use of standalone agents, and making sure the important data surfaces first. It’s more about attention and getting the RIGHT context in just in time. It’s like a cache… When you get it right it feels magic.
Short version IMHO we need to get what we put in the prompt solid before we start swatting ents out.
Condense through summary before even getting to RAG… I’m doing complex automation and about 800 ents. No true rag yet - just tools and state instead it’s a lot of strategically placed breadcrumbs and good solid instructions.
Note I didn’t say no rag. Just, I’m convinced that performance comes from getting tight in prompts tools and instructions and we can do a lot of that before RAG. If we start removing ents from context then we lose a lot of the. ‘oh i can do that now to live context of the home.’
Where I see RAG shining here (at least for HAs use case in almost all cases) is presummary… We’re seriously under using this especially with local llm. Local pulls ahead of the cloud when you can keep a pipe full of small jobs. Using that RAG you can get deep rich summary context compiled just before it’s consumed. If yoj look hard at Fridays architecture you’ll see RAG there in two stages… (The monastery is the LLMs door to the RAG enabled agent) Cabinets are a kind of mini-rag but more important it’s structured, addressable parking for the presummaries that the llm can consume on demand and it allows the llm to act like a person in the home in a way that feels natural. (ultimately they will be two views and access modes of the same data)
You can also share knowledge between multiple specialists to pull heavy context out of the prompt… Then, you get lean n mean in the Frontline llm but rich context and tools to reach into. - saving space for ents and live context.
100% for context compression. Not sure about that exact method. But let’s figure it out.